Kentino AI Node A100 — 8× NVIDIA A100 SXM4 80 GB, 640 GB HBM2e
Enterprise-grade platform, assembled and tested in the EU.
Kentino AI Node A100
640 GB HBM2e, NVSwitch-Pooled
8x A100 SXM4 80 GB | 4,992 INT8 TOPS
The only node tier with a sane cost per gigabyte of VRAM.
An 8-GPU node carrying eight NVIDIA A100 SXM4 80 GB modules on an NVLink/NVSwitch baseboard: 640 GB of HBM2e at 2.04 TB/s per GPU with 600 GB/s any-to-any GPU bandwidth. Ampere is two generations back, and for a large class of work that is exactly why you want it - the same 640 GB pool as a Hopper H100 node, at substantially less money and 2.4 kW less heat. If your models run at bf16 or Q4 and you are not going to use fp8, you should not be paying for it.
Hardware
| Component | Detail |
|---|---|
| GPUs | 8x NVIDIA A100 SXM4 80 GB HBM2e (Ampere, 400 W per module) |
| VRAM pool | 640 GB HBM2e, NVSwitch-pooled |
| Memory bandwidth | 2.04 TB/s per GPU |
| GPU interconnect | NVLink 3 via NVSwitch, 600 GB/s any-to-any |
| Compute | 2,496 TFLOPS bf16 dense | 4,992 INT8 TOPS. No fp8 tensor path |
| CPUs | 2x AMD EPYC 7643 (Milan, 48C/96T each) on the Supermicro AS-4124GO-NART+ platform |
| System memory | 1 TB DDR4-3200 ECC RDIMM (16x 64 GB) |
| Storage | 2x 1.92 TB U.2 NVMe boot + 2x 3.84 TB U.2 NVMe data |
| Networking | 2x NVIDIA ConnectX-7 400 Gb/s |
| Chassis | 4U rack-mount, front-to-back airflow, 24/7 rated |
| Power | 4x 3000 W redundant |
What it runs
640 GB at bf16 and Q4 covers most production serving and fine-tuning. Estimates are for this 8-GPU configuration with tensor parallelism.
| Component | Detail |
|---|---|
| Llama 3.3 70B / Qwen3-72B | bf16 full precision. ~35-50 tok/s single stream, 1,500+ tok/s batched |
| Qwen3-235B-A22B MoE | Q4 across the full 640 GB pool |
| Mistral Large 2 123B | Q4-Q8 with context headroom |
| QLoRA / full fine-tune 70B | 8-way tensor parallel, NVSwitch-backed |
| Classic HPC | 9.7 TFLOPS FP64 per GPU - 77.6 TFLOPS FP64 per node |
| Multi-tenant serving | several 30B-70B models resident at once |
Throughput figures are estimates derived from published vendor specifications and community benchmarks for this GPU generation, not measurements taken on this exact configuration. We will give you measured numbers for your model and context length before you commit.
Power and cooling
Roughly 4.3 kW sustained: 3.2 kW of GPU plus CPU, fans and storage. That is 2.4 kW less than a Hopper node and it is the practical selling point - 4.3 kW fits a conventional datacentre rack with front-to-back airflow and no aisle containment retrofit. Over three years at EU industrial electricity rates the difference against an H100 node is a real line in the TCO, not a rounding error.
What you give up, and export control
No fp8. Ampere predates the format. Models shipping fp8-native - the DeepSeek-V3 family, Hunyuan-Large - run here at bf16 or Q4 instead, which costs VRAM and throughput. If your roadmap is fp8 checkpoints, buy Hopper. This is the single real limitation of the tier and we would rather you hear it now.
Export control. A100 80 GB is above the Advanced Computing Item thresholds (ECCN 3A090.a), so it needs the same KYC form and End User Certificate as the Hopper nodes.
Late-life hardware. Ampere is near end of channel life. Warranty horizon and spares availability are shorter than on current generation, and we will tell you exactly what we can commit to in writing before you order.
What you get from Kentino
Kentino supplies, configures and supports this node as a complete deployment: rack and power design against your site's actual feed, front-to-back airflow sizing for 24/7 operation, OS and driver stack (Ubuntu with pinned NVIDIA driver, CUDA, container toolkit), the inference or training stack of your choice, burn-in with a written report, EU-based logistics and warranty.
We also handle the export paperwork with you rather than mailing you a form and hoping. That part takes longer than the hardware and is where most first-time buyers of this class of machine get stuck.
Price shown is EUR excluding VAT, EXW Prague (Incoterms 2020). Collection, freight, insurance and any destination import duties are for the buyer’s account. Lead time on request.
Tell us the models you need to serve and we will tell you honestly whether this node or a Hopper node is the right purchase - the answer is genuinely sometimes this one.
The questions buyers ask us most often before ordering a server.
How long does it take?
Machines built from components we hold ship quickly; anything requiring a specific GPU generation depends on supply. We give you a date before you pay, and if it moves we tell you rather than letting you find out.
Can the configuration be changed before you build it?
Almost always. GPUs, memory, storage and cooling are chosen per order, and the listed configuration is a starting point rather than a fixed package. If you need more VRAM, faster storage or a different cooling approach, say so before you order and we will quote the change.
Can I collect the server in person?
You can. Our warehouse is in Prague, and collection in person is welcome — most people who come use the visit to go through the machine with our engineer and ask the questions that are awkward over email. For orders within the Czech Republic we also try to deliver personally and walk you through the setup on site.
Can I talk to someone who actually understands the workload?
Yes. We have an engineer who works on AI systems specifically, not a general sales desk. If your question is about batch sizes, quantisation, interconnect or where your bottleneck will be, ask it — that conversation usually changes the configuration for the better.
Which model can I run on this configuration?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
How do you test a server before shipping?
We run it against actual AI workloads rather than synthetic scores: inference throughput, sustained load behaviour and thermals under continuous operation. You get the benchmark results with the machine, so the performance you were promised is the performance you can verify on day one.
Is the server ready to run when it arrives?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
Ships from our EU warehouse. Heavy items may require freight arrangement — contact us for a shipping quote and lead time. 2-year limited warranty with advanced RMA support; extended warranty available.
Not exactly what you need?
Tell us your workload and we'll spec this platform around it — GPUs, memory, storage and cooling matched to what you actually run.